DeepSeek V4 Pro: Self-Hosting & Deployment Guide
Up to 10%DeepSeek V4 Pro is DeepSeek’s flagship open-weights Mixture-of-Experts language model with a 1 million token context window and strong reasoning and coding capabilities.
What is DeepSeek V4 Pro?
DeepSeek V4 Pro is a 1.6-trillion-parameter Mixture-of-Experts large language model from DeepSeek with around 49 billion activated parameters and a 1 million token context window. It is mainly used for advanced reasoning tasks such as complex problem solving, long-horizon agent workflows, and high-end software engineering and coding assistance. It is also used for long-context analysis, knowledge-intensive question answering, and tool-using applications that require function calling and structured outputs. It belongs to the DeepSeek V4 family and succeeds earlier DeepSeek models such as DeepSeek-R1 and prior V-series models.
How to run DeepSeek V4 Pro yourself
Open weights (mit), 1598.8B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.
DeepSeek V4 Pro VRAM requirements
| Precision / quantization | Weights | Figure | Source |
|---|---|---|---|
| BF16 (full precision) | 3197.7 GB | Estimate | 1598.8B params × 2 bytes |
| FP8 | 1606.5 GB | File size | sgl-project/DeepSeek-V4-Pro-FP8 |
File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.
GPU requirements: deploy DeepSeek V4 Pro on your hardware
| Hardware | BF16 (full precision) | FP8 |
|---|---|---|
| NVIDIA H200141 GB | — | — |
| NVIDIA H10080 GB | — | — |
| NVIDIA A100 80GB80 GB | — | — |
| NVIDIA L40S48 GB | — | — |
GPUs needed per variant.
| Hardware | BF16 (full precision) | FP8 |
|---|---|---|
| AWS p5.48xlarge8× H100 · 640 GB | — | — |
| AWS p4de.24xlarge8× A100 80GB · 640 GB | — | — |
| AWS g6e.12xlarge4× L40S · 192 GB | — | — |
| Google Cloud a3-highgpu-8g8× H100 · 640 GB | — | — |
| Azure ND H100 v58× H100 · 640 GB | — | — |
Whole-instance GPU memory; public instance specs.
Serve it yourself
pip install -U vllm
vllm serve deepseek-ai/DeepSeek-V4-Propip install 'sglang[all]'
python -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V4-ProCommands use the official repositories above; vLLM and SGLang expose an OpenAI-compatible endpoint.
Skip the deploy — use LLM.API as your endpoint
No GPUs, no quantization trade-offs. Same model, OpenAI-compatible, up to 10% below list price.
Providers
Different companies host the same model. OpenRouter routes your request to one of them based on the routing mode you pick — Balanced (price + speed), Nitro (fastest), or Exacto (highest tool-calling accuracy).
List price by provider
InputOutputProvider list prices; the LLM.API discount applies on top.
| Provider | Input | Output | Cache read /M | Latency | Throughput | Uptime |
|---|---|---|---|---|---|---|
| in $1.74; out $3.48 per 1M tokens | 1M tokens | tools, streaming, reasoning, web search, JSON, structured | ||||
| in $1.32; out $3.96 per 1M tokens | 1M tokens | tools, streaming, reasoning, web search, JSON, structured | ||||
| in $1.6; out $3.2 per 1M tokens | 1M tokens | tools, streaming, reasoning, web search, JSON, structured | ||||
| in $1.32; out $3.96 per 1M tokens | 1M tokens | tools, streaming, reasoning, web search, JSON, structured |
Prices and availability from the LLMAPI catalogue, updated nightly. Last updated 11 Oct 2026.
Try this model
Test DeepSeek V4 Pro right here — free to start.
Suggestions for your first prompt
Code snippet
Call the model through the OpenAI-compatible API.
Switching from OpenAI? Change 2 lines.
- base_url="https://api.openai.com/v1"
- api_key="YOUR_OPENAI_KEY"
+ base_url="https://api.llmapi.ai/v1"
+ api_key="YOUR_LLMAPI_KEY"Everything else stays the same — SDK, message format, tools, streaming. Set model to deepseek-v4-pro and the call runs unchanged.
curl https://api.llmapi.ai/v1/chat/completions \
-H "Authorization: Bearer $LLMAPI_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-pro",
"messages": [
{"role": "user", "content": "Give me three crisp launch checklist items."}
],
"reasoning_effort": "high"
}'from openai import OpenAI
client = OpenAI(
api_key="YOUR_LLMAPI_KEY",
base_url="https://api.llmapi.ai/v1",
)
resp = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[
{"role": "system", "content": "You are a precise product assistant."},
{"role": "user", "content": "Give me three crisp launch checklist items."},
],
reasoning_effort="high",
)
print(resp.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.LLMAPI_KEY,
baseURL: "https://api.llmapi.ai/v1",
});
const resp = await client.chat.completions.create({
model: "deepseek-v4-pro",
messages: [
{ role: "user", content: "Give me three crisp launch checklist items." },
],
});
console.log(resp.choices[0].message.content);stream = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Write a haiku about latency."}],
stream=True,
)
for chunk in stream:
delta = chunk.choices[0].delta.content
if delta:
print(delta, end="", flush=True){
"model": "deepseek-v4-pro",
"messages": [
{"role": "system", "content": "You are a precise product assistant."},
{"role": "user", "content": "Give me three crisp launch checklist items."}
],
"reasoning_effort": "high"
}ENDPOINTS
Supported endpoints & features
Base URL https://api.llmapi.ai/v1 · model ID deepseek-v4-pro · OpenAI-compatible.
POST /v1/chat/completionsChat Completions — supported
| Feature | Status | How to set it |
|---|---|---|
| Function calling (tools) | Supported | tools, tool_choice |
| Parallel tool calls | Not supported | parallel_tool_calls |
| JSON mode | Supported | response_format: {"type": "json_object"} |
| Structured outputs | Supported | response_format: {"type": "json_schema", …} |
| Streaming | Supported | stream: true |
| Vision (image input) | Not supported | image_url content parts |
| Web search | Supported | web_search |
| Reasoning effort | high · max | reasoning_effort: "high" |
ERRORS
Errors & fallback routing
| Code | What it means | What to do |
|---|---|---|
401 | Missing or invalid API key. | Send Authorization: Bearer YOUR_LLMAPI_KEY and check the key is active. |
429 | Rate limited, or the upstream provider is throttling the request. | Back off and retry with jitter; the gateway also retries the request on another provider where one is available. |
5xx | Upstream provider error or timeout. | Retry; fallback routing sends the retry to the next healthy provider for this model. |
When a provider fails or throttles, the request is routed to the next provider serving this model — the ones listed in the providers table above.
DeepSeek V4 Pro: Self-Hosting & Deployment Guide in production
What the model is good at.
-
Advanced Chat
Engages in multi-turn conversations, follows complex instructions, and maintains context across long interactions for diverse assistant-style tasks.
-
Image Understanding
Analyzes input images, recognizing objects and visual details to support tasks like description, comparison, and visual reasoning.
-
Code Monitoring
Supports reviewing and reasoning about code or logs, helping detect issues, explain behavior, and guide debugging steps.
-
Multilingual Translation
Translates between multiple languages, preserving key meaning and style for everyday text and technical content.
-
Text Recognition
Extracts and interprets textual content from provided images, enabling downstream understanding, search, and transformation of visual documents.
What you get when you call DeepSeek V4 Pro: Self-Hosting & Deployment Guide through LLM.API.
-
Unified AI Routing
Automatically route each request to the best model across providers based on latency, cost, and quality—without changing your integration or redeploying.
One endpoint. Any model. -
Cost-Aware Orchestration
Control spend with configurable policies that downshift to cheaper models when possible and reserve premium models only where they truly matter.
Optimize quality per dollar. -
Resilient Fallback Logic
Eliminate single-provider downtime with automatic fallbacks that retry on alternate models and regions while preserving request shape and semantics.
Stay up when APIs fail. -
Deep LLM Observability
Get full visibility into tokens, latency, errors, and provider health with request-level traces that plug into your existing monitoring stack.
See every token hop. -
Task-Level Abstractions
Define tasks like chat, tools, or RAG once and let LLM.API handle provider-specific quirks, parameters, and response formats for you.
Code to tasks, not vendors. -
High-Throughput Batch Jobs
Run massive inference and evaluation workloads with parallelized, rate-safe batching that maximizes throughput across providers without throttling or manual sharding.
Scale jobs, not scripts.
When to Use — When NOT to Use
Use it if...
- You need a cost-effective, general-purpose LLM for a wide range of tasks.
- You need strong multilingual understanding and generation across many major world languages.
- Your use case involves complex reasoning or coding that benefits from a powerful frontier model.
- You need good performance on math, logic, and structured problem-solving without frontier-model pricing.
- Your use case involves building chatbots, agents, or tools needing tool-use and web-calling abilities.
- You need an alternative to US-based providers for redundancy, jurisdiction, or data-governance reasons.
Avoid if...
- You need guaranteed access to US or EU enterprise-grade compliance, certifications, and legal guarantees.
- Your workload requires tight integration with the OpenAI ecosystem or proprietary OpenAI-specific features.
- You need heavily audited safety filters and mature governance comparable to top US hyperscale providers.
- Your workload requires extremely low latency from US data centers with strict geographic residency.
- You need battle-tested reliability under massive global production scale with long historical uptime records.
- Your workload requires fully transparent, extensively documented training data sources meeting strict compliance rules.
COMMUNITY
What developers say about DeepSeek V4 Pro
Summarised from public developer discussions and write-ups. Opinions are the sources’, not LLM.API’s.
- Developers on Reddit and Hugging Face report disappointment regarding the lack of native multimodality, noting that the model lacks the visual reasoning and multimodal output capabilities found in competitors like Qwen or Kimi.
- Engineers at inDrive found the model highly effective for coding tasks, achieving 80.6% on SWE-bench, though they noted it is significantly more verbose than Claude and occasionally stumbles on complex agentic workflows.
- Developers indicate that while the model is open-source, the 1.6T parameter size (roughly 862 GB) makes local hosting impossible for consumer hardware, requiring datacenter-grade GPU clusters like 8x H100s for inference.
SOURCES
Frequently Asked Questions
-
What is DeepSeek V4 Pro?
DeepSeek V4 Pro is a large language model by DeepSeek focused on strong reasoning, coding, and general-purpose text generation.
-
What modalities does DeepSeek V4 Pro support via LLM.API?
DeepSeek V4 Pro is available as a text-only model on LLM.API, accepting text prompts and returning text completions or chat responses.
-
How is DeepSeek V4 Pro typically priced on LLM.API?
DeepSeek V4 Pro is billed on a pay-as-you-go basis per thousand input and output tokens, with exact rates shown in your LLM.API pricing dashboard.
-
What is the context window of DeepSeek V4 Pro?
DeepSeek V4 Pro supports a large-context window suitable for long conversations and multi-file coding tasks; check LLM.API docs for the current token limit.
-
How fast is DeepSeek V4 Pro in terms of latency?
DeepSeek V4 Pro generally returns first tokens within a few seconds, with total latency depending on prompt size, response length, and LLM.API load.
-
What is DeepSeek V4 Pro best suited for?
DeepSeek V4 Pro is best for complex reasoning, code generation and debugging, data analysis assistance, and high-quality general-purpose writing.
-
How do I call DeepSeek V4 Pro through LLM.API?
You select the DeepSeek V4 Pro model name in your LLM.API request payload, pass your prompt as messages or text, and authenticate with your API key.
-
How does DeepSeek V4 Pro compare to similar frontier models?
DeepSeek V4 Pro offers competitive reasoning and coding quality, often at a lower token cost than many frontier models from larger providers.
-
What are the main limitations of DeepSeek V4 Pro?
DeepSeek V4 Pro can still hallucinate, may lack very recent knowledge, and should not be trusted alone for high-stakes legal, financial, or medical decisions.
-
Does DeepSeek V4 Pro support streaming responses on LLM.API?
Yes, DeepSeek V4 Pro can stream tokens incrementally when you enable streaming in your LLM.API request, reducing perceived latency for long outputs.
COMPARE
Competitive Models
Gemini 3.1 Pro (Preview)
$2 in / $12 out per 1M tokens, 1,048,576 context — the closest current alternative to DeepSeek V4 Pro.
GPT-5.6 Terra
$2 in / $12 out per 1M tokens, 1,050,000 context — the closest current alternative to DeepSeek V4 Pro.
Qwen 3.8 Max
$2 in / $6 out per 1M tokens, 1,000,000 context — the closest current alternative to DeepSeek V4 Pro.
DeepSeek V3.2 is a large open-source Mixture-of-Experts language model from DeepSeek that emphasizes high reasoning performance and efficient long‑context inference. It is notable for its DeepSeek Sparse Attention and multi-latent attention mechanisms, which significantly cut compute and memory costs for long sequences.
DeepSeek V4 Flash
Up to 30%DeepSeek V4 Flash is an efficiency-optimized Mixture-of-Experts large language model from DeepSeek, featuring a 1M-token context window and fast inference for high-throughput applications.
DeepSeek V4 Flash (free)
Up to 30%DeepSeek V4 Flash (free) is an open-source, efficiency-optimized Mixture-of-Experts language model from DeepSeek, offering a 1M-token context window with only 13B parameters activated per token out of 284B total. It is designed to deliver fast, cost-effective long-context reasoning, coding, and agentic workflows.
Get one key to every model
Swap your API key. Keep your code.